C++ Object Persistence with ODB

Copyright © 2009-2010 Code Synthesis Tools CC

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, version 1.3; with no Invariant Sections, no Front-Cover Texts and no Back-Cover Texts.

Revision 1.0, September 2010

This revision of the manual describes ODB 1.0.0 and is available in the following formats: XHTML, PDF, and PostScript.

Table of Contents

2 Hello World Example

In this chapter we will examine how to create a simple C++ application that relies on ODB for object persistence using the traditional "Hello World" example. In particular, we will discuss how to declare persistent classes, generate database support code, as well as compile and run our application. We will also learn how to make objects persistent as well as query, update and delete persistent objects.

The code presented in this chapter is based on the hello example which can be found in the odb-examples package of the ODB distribution.

2.1 Declaring a Persistent Class

In our "Hello World" example we will depart slighly from the norm and say hello to people instead of the world. People in our application will be represented as objects of C++ class person which is saved in person.hxx:

// person.hxx
//

#include <string>

class person
{
public:
  person (const std::string& first,
          const std::string& last,
          unsigned short age);

  const std::string&
  first () const;

  const std::string&
  last () const;

  unsigned short
  age () const;

  void
  age (unsigned short);

private:
  std::string first_;
  std::string last_;
  unsigned short age_;
};

In order not to miss anyone whom we need to greet, we would like to save the person objects in a database. To achive this we declare the person class as persistent:

// person.hxx
//

#include <string>

#include <odb/code.hxx>     // (1)

#pragma db object           // (2)
class person
{
  ...

private:
  person () {}              // (3)

  friend class odb::access; // (4)

  #pragma db id auto        // (5)
  unsigned long id_;        // (5)

  std::string first_;
  std::string last_;
  unsigned short age_;
};

To be able to save person objects in the database we had to make five changes, marked with (1) to (5), to the orignal class definition. The first change is the inclusion of the ODB headers core.hxx. This headers provides a number of core ODB declarations, such as odb::access, that are used to define peristent classes.

The second change is the addition of db object pragma just before the class definition. This pragma tells the ODB compiler that the class that follows is persistent. Note that making a class persistent does not mean that all objects of this class will automatiacally be stored in the database. You would still create ordinary or transient instances of this class just as you would before. The difference is that now you can make such transient instances persistent, as we will see shortly.

The third change is the addition of the default constructor. The ODB-generated database support code will use this constructor when instantiating an object from the persistent state. As we have done for the person class, you can make the default constructor private or protected if you don't want to make it available to the ordinary users of your class.

With the fourth change we make the odb::access class friend of our person class. This is necessary to make the default constructor and the data members accessible to the ODB support code. If your class has public default constructor and public data members, then the friend declaration is unnecessary.

The final change adds a data member called id_ which is preceded by another pragma. In ODB every persistent object must have a unique, within its class, identifier. Or, in other words, no two persistent instances of the same type have equal identifiers. For our class we use an integer id. The db id auto pragma that preceeds the id_ member tells the ODB compiler that the following member is the object's id. The auto specifier indicates that it is a database-assigned id. A unique id will be automatically generated by the database and assigned to the object when it is made persistent.

In this example we choose to add an identifier because none of the existing members could serve the same purpose. However, if a class already has a member with suitable properties, then it is natural to use that member for an identifier. For example, if our person class contained some form of personal identification (SSN in the United States or ID/passport number in other countries), then we could use that as an id. Or, if we stored an email associated with each person, then we could have used that since each person is presumed to have a unqiue email address:

class person
{
  ...

  #pragma db id
  std::string email_;

  std::string first_;
  std::string last_;
  unsigned short age_;
};

Now that we have the header file with the persistent class, let's see how to generate that database support code that we talked about.

2.2 Generating Database Support Code

The persistent class definition that we created in the previous section was particularly light on code that could actualy do the job and store the person't data to a database. There was no serialization or deserialization code, not even data member registration, that you would normally have to write by hand in other ORM libraries for C++. This is because in ODB code that translates between the database and C++ representations of an object is automatically generated by the ODB compiler.

To compile the person.hxx header we created in the previous section and generate the support code for the MySQL database we invoke the ODB compiler from a terminal (UNIX) or a command prompt (Windows):

odb -d mysql --generate-query person.hxx

We will use MySQL in the reminder of this chapter though other supported database systems can be used instead.

If you haven't installed the common ODB runtime library (libodb) or installed it into a directory where the C++ compiler doesn't search for headers by default, then you may get the following error:

person.hxx:10:24: fatal error: odb/core.hxx: No such file or directory

To resolve this you will need to specify libodb headers location with the -I preprocessor option, for example:

odb -I.../libodb -d mysql --generate-query person.hxx

Here .../libodb represents the path to the libodb directory.

The above invocation of the ODB compiler produces three C++ files: person-odb.hxx, person-odb.ixx, person-odb.cxx. You normally don't use types or functions contained in these files directly. Rather, all you have to do is include person-odb.hxx in C++ files where you are performing database operations with classes from person.hxx as well as compile person-odb.cxx and link the resulting object file to your application.

You may be wondering what is the --generate-query option for. It instructs the ODB compiler to generate optional query support code that we will use later in our "Hello World" example. Another option that we will find useful is --generate-schema. This option makes the ODB compiler generate a fourth file, person.sql, which contains the database schema for the classes defined in person.hxx:

odb -d mysql --generate-query --generate-schema person.hxx

If you would like to see the list of all the available options, refer to the ODB Compiler Command Line Manual.

Now that we have the persistent class and the database support code, the only part that is left is the application code that does something useful with all this. But before we move on to the fun part, let first learn how to build and run an application that uses ODB. This way when we have some application code to try, there are no more delays before we can run it.

2.3 Compiling and Running

Assuming that the main() function with some application code is saved in driver.cxx and the database support code and schema are generated as described in the previous section, to build our application we will first need to compile all the C++ source files and then link them with two ODB runtime libraries.

On UNIX, the compilation part can be done with the following commands (for Microsoft Visual Studio setup, see the odb-examples package):

c++ -c driver.cxx
c++ -c person-odb.cxx

Similar to the ODB compilation, if you get an error stating that a headers in odb/ or odb/mysql directory in not found. In this case you will need to use the -I preprocessor option to specify the location of the common ODB runtime library (libodb) and MySQL ODB runtime library (libodb-mysql).

Once the compilation is done, we can link the application with the following command:

c++ -o driver driver.o person-odb.o -lodb-mysql -lodb

Notice that we link our application with two ODB libraries: libodb which is a common runtime library and libodb-mysql which is a MySQL runtime library (if you use another database, then the name of this library will change accordingly). If you get an error saying that one of these libraries could not be found, then you will need to use the -L linker option to specify their locations.

Before we can run our application we need to create a database schema using the generated person.sql file. For MySQL we can use the mysql client program, for example:

mysql --user=odb_test --database=odb_test < person.sql

The above command will login to a local MySQL server as user odb_test without a password and use database named odb_test. Note that after executing this command all data stored in the odb_test database will be deleted.

Once the database schema is ready, we run our application using the same login and database name:

./driver --user odb_test --database odb_test

2.4 Making Objects Persistent

Now that we have the infrastructure work out of the way, it is time to see our first code fragment that interracts with the database. In this section we will learn how to make person objects persistent:

// driver.cxx
//

#include <memory>   // std::auto_ptr
#include <iostream>

#include <odb/database.hxx>
#include <odb/transaction.hxx>

#include <odb/mysql/database.hxx>

#include "person.hxx"
#include "person-odb.hxx"

using namespace std;
using namespace odb;

int
main (int argc, char* argv[])
{
  try
  {
    auto_ptr<database> db (new mysql::database (argc, argv));

    unsigned long john_id, jane_id, joe_id;

    // Create a few persistent person objects.
    //
    {
      person john ("John", "Doe", 33);
      person jane ("Jane", "Doe", 32);
      person joe ("Joe", "Dirt", 30);

      transaction t (db->begin_transaction ());

      db->persist (john);
      db->persist (jane);
      db->persist (joe);

      t.commit ();

      // Save object ids for later use.
      //
      john_id = john.id ();
      jane_id = jane.id ();
      joe_id = joe.id ();
    }
  }
  catch (const odb::exception& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}

Let's examine this code piece by piece. At the beginnig we include a bunch of headers. Those include odb/database.hxx and odb/transaction.hxx which define database system-independant odb::database and odb::transaction interfaces. Then we include odb/mysql/database.hxx which defines the MySQL implementation of the database interface. Finaly, we include person.hxx and person-odb.hxx which define our persistent person class.

Once we are in main(), the first thing we do is create the MySQL database object. Notice that this is the last line in driver.cxx that mentions MySQL explicitly; the rest of the code works though the common interfaces and is database system-independant. We use the argc/argv mysql::database constructor which automatically extract the database parameters, such as login name, passowrd, database name, etc., from the command line. In your own applications you may prefer to use other versions of the mysql::database constructors which allow you to pass this information directly (@@ ref MySQL database).

Next we create three person objects. Right now they are transient objects, which means that if we terminate the application at this point, they will be gone without any evidence of them ever existed. The next line starts a database transaction. We discuss transactions in detail later in this manual. For now all we need to know is that all ODB database operations must be performed within a transaction and that a transaction is an atomic unit of work; all database operations performed within a transaction either succeed (commited) together or are automatically undone (rolled back).

Once we are in a transaction, we call the persist() database function on each of our person objects. At this point the state of each object is saved in the database. However, note that this state is not permanent until and unless the transaction is commited. If, for example, our application crashes at this point, there will still be no evidence of our objects ever existed.

In our case one more thing happens when we call persist() on a person object. Remember that we decided to use database-assigned identifiers for our objects. The call to persist() is where this assignment happens. Once this function returns, the id_ member contains this object's unique identifier.

After we have persisted our objects, it is time to commit the transaction and make the changes permanent. Only after the commit() function returns succefully are we guaranteed that the objects are made persistent. Following the crashing example, if our application terminates after the commit for whatever reason, the objects' state in the database will remain intact. In fact, as we will discover shortly, our application can be restarted and load the orignal objects from the database. Note also that a transaction must be commited explicitly with the commit() call. If the transaction object leaves scope without the transaction beeing explicitly commited or rolled back, it will be automatically rolled back. This behavior allows you not to worry about exceptions being thrown within a transaction; if they cross the transaction boundaries, the transaction will be automatically rolled back and all the changes made to the database undone.

After the transaction has been commited, we save the persistent objects' ids in local variables. We will use them later in this chapter to perform other database operations on our persistent objects. You might have noticed that our person class doesn't have the id() function that we use here. To make our code work we need to add a simple accessor with this name that returns the value of the id_ data member.

The final bit of code in our example is the catch block that handles the ODB exceptions. We do this by catching the base ODB exception and printing the diagnostics. (@@ Ref exceptions)

Let's now compile (see @@ Ref "Compiling and Running") and then run our first ODB application:

mysql --user=odb_test --database=odb_test < person.sql
./driver --user odb_test --database odb_test

Our first application doesn't print anything except for error messages so we can't really tell whether it actually stored the objects' state in the database. While we will extend our application to be more enternaining, for now we can use the mysql client to examine the database content. It will also give us a feel for how the object are stored:

mysql --user=odb_test --database=odb_test

Welcome to the MySQL monitor.

mysql> select * from person;

+----+-------+------+-----+
| id | first | last | age |
+----+-------+------+-----+
|  1 | John  | Doe  |  33 |
|  2 | Jane  | Doe  |  32 |
|  3 | Joe   | Dirt |  30 |
+----+-------+------+-----+
3 rows in set (0.00 sec)

mysql> quit

In the next section we will examine how to query persistent objects from our application.

2.4 Querying Persistent Objects

So far our application doesn't resemble a typical "Hello World" example. It doesn't print anything except for error messages. Let's change that and teach our application to say hello to people from our database. To make it a bit more interesting, let's say hello only to people over 30:

// driver.cxx
//

...

int
main (int argc, char* argv[])
{
  try
  {
    ...

    // Create a few persistent person objects.
    //
    {
      ...
    }

    typedef odb::query<person> query;
    typedef odb::result<person> result;

    // Say hello to those over 30.
    //
    {
      transaction t (db->begin_transaction ());

      result r (db->query<person> (query::age > 30));

      for (result::iterator i (r.begin ()); i != r.end (); ++i)
      {
        cout << "Hello, " << i->first () << "!" << endl;
      }

      t.commit ();
    }
  }
  catch (const odb::exception& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}

The first half of our application is the same as before and is replaced with "..." in the above listing for brievety. Again, let's examine the rest of it piece by piece.

The two typedefs create convenient aliases for two template instantiations that will be used a lot in our application. The first is the query type for the person objects and the second is the result type of that query.

Then we begin a new transaction and call the query() database function. We pass a query expression (query::age > 30) which limits the returned objects only to those with age greater than 30. We also save the result of the query in a local variable.

The next few lines perform a pretty standard for-loop iteration over the result sequence printing hello for every returned person. Then we commit the transaction and we are node. Let's see what this application will print:

mysql --user=odb_test --database=odb_test < person.sql
./driver --user odb_test --database odb_test

Hello, John!
Hello, Jane!

That looks about right but how do we know that the query actually used the database instead of just using some in-memory artifacts of the earlier persist() calls. One way to test this would be to comment out the first transaction in our application and re-run it without re-creating the database schema so that the objects that were persisted during the previous run will be returned. Alternatively, we can just re-run the same application without re-creating the schema and notice that we now how duplicate objects:

./driver --user odb_test --database odb_test

Hello, John!
Hello, Jane!
Hello, John!
Hello, Jane!

What happens here is that the previous run of our application persisted a set of person objects and when we re-run the application, we persist another set with the same names but with different id. When we later run the query, matches from both sets are returned. We can change the line where we print the "Hello" string as follows to illustrate this point:

cout << "Hello, " << i->first () << " (" << i->id () << ")!" << endl;

If we now re-run this modified program, we will get the following output:

./driver --user odb_test --database odb_test

Hello, John (1)!
Hello, Jane (2)!
Hello, John (4)!
Hello, Jane (5)!
Hello, John (7)!
Hello, Jane (8)!

The identifiers 3, 6, and 9 that miss from the above list belong to the "Joe Dirt" objects which are not selected by this query.

2.5 Updating Persistent Objects

While making objects persistent and then querying them are useful oprations, most applications will also need to change the object's state and then make these changes persistent. Let's illustrate this by updating Joe's age who just had a birthday:

// driver.cxx
//

...

int
main (int argc, char* argv[])
{
  try
  {
    ...

    unsigned long john_id, jane_id, joe_id;

    // Create a few persistent person objects.
    //
    {
      ...

      // Save object ids for later use.
      //
      john_id = john.id ();
      jane_id = jane.id ();
      joe_id = joe.id ();
    }

    // Joe Dirt just had a birthday, so update his age.
    //
    {
      transaction t (db->begin_transaction ());

      auto_ptr<person> joe (db->load<person> (joe_id));
      joe->age (joe->age () + 1);
      db->store (*joe);

      t.commit ();
    }

    // Say hello to those over 30.
    //
    {
      ...
    }
  }
  catch (const odb::exception& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}

The beginning and the end of this transaction are the same as the previous two. Once within a transaction, we call the load() database function to instantiate a person object with Joe's persistent state. We pass Joe's object identifer that we stored earlier when we made this object persistent.

With the instantiated object in hand we increment the age and call the store() database function to update the object's state in the database. Once the transaction is commited, the changes are made permanent in the database.

If we now run this application, we will see Joe in the output since he is now over 30:

mysql --user=odb_test --database=odb_test < person.sql
./driver --user odb_test --database odb_test

Hello, John!
Hello, Jane!
Hello, Joe!

What if we didn't have an identifier for Joe? Maybe this object was made persisted in another run of our application or by another application altogether. Provided that we have only one Joe Dirt in the database, we can use query to come up with an alternative implementation of the above transaction:

    // Joe Dirt just had a birthday, so update his age. An
    // alternative implementation without using the object id.
    //
    {
      transaction t (db->begin_transaction ());

      result r (db->query<person> (query::first == "Joe" &&
                                   query::last == "Dirt"));

      result::iterator i (r.begin ());

      if (i != r.end ())
      {
        auto_ptr<person> joe (*i);
        joe->age (joe->age () + 1);
        db->store (*joe);
      }

      t.commit ();
    }

2.5 Deleting Persistent Objects

The last operation that we will discuss in this chapter is deleting the persistent object from the database. The following code fragment shows how we can delete an object given its identifier:

    // John Doe is no longer in our database.
    //
    {
      transaction t (db->begin_transaction ());
      db->erase<person> (john_id);
      t.commit ();
    }

To delete John from the database we start a transaction, call the erase() database function with John's object id, and commit the transaction. After the transaction is commited the erased object is no longer persistent.

If we don't have an object id handy, we can use query to find and delete the object:

    // John Doe is no longer in our database. An alternative
    // implementation without using the object id.
    //
    {
      transaction t (db->begin_transaction ());

      result r (db->query<person> (query::first == "John" &&
                                   query::last == "Doe"));

      result::iterator i (r.begin ());

      if (i != r.end ())
      {
        auto_ptr<person> john (*i);
        db->erase (*john);
      }

      t.commit ();
    }

2.5 Summary

This chapter presented a very simple application which, nevertheless, excercised all core database functions: persist(), query(), load(), store(), and erase(). We also saw that writing an application that uses ODB involves the following steps:

  1. Declare persistent classes in header files.
  2. Compile these headers to generate database support code.
  3. Link the application with the support code and two ODB runtime libraries.

Do not be concerned if, at this point, much appears unclear. The intent of this chapter is to give you only a general idea of how to persist C++ objects with ODB. We will cover all the details throughout the remainder of this manual.

3 Working Title

3.1 Base Concepts

The term database can refer to three distinct things: a general notion of a place where an application stores its data, a software implementation for managing this data (for example MySQL), and, finally, some database software implementations may manage several data stores which are usually distinguished by name. This name is also commonly referred to as database.

In this manual, when we use just the word database, we refer to the first meaning above, for example, "The store() function saves the object's state to the database." The term Database Management System (DBMS) is often used to refer to the second meaning of the words database. In this manual we will use the term database system for short, for example, "Database system-independant application code." Finally, to distinguish the third meaning from the other two we will use the term database name, for example, "The second option specfies the database name that the application should use to store its data."

In C++ there is only one notion of a type and an instance of a type. For example, a fundamental type, such as int, is, for the most part, treated the same as a user defined class type. However, when it comes to persistence, we have to place certain restrictions and requirements on certain C++ types that can be stored in the database. As a result, we devide persistent C++ types into two groups: object types and value types. An stances of an object type is called an object and an instance of a value type — a value.

An object is an independant entity. It can be stored, updated, and deleted in the database independant of other objects or values. An object has an identifier, called object id, that is unique among all instances of an object type within a database. An object consits of data members which are either values or references to other objects. In contrast, a value can only be stored in the database as part of an object and doesn't have its own unique identifier.

An object type is a C++ class. Because of this one to one relationship, we will use terms object type and object class interchangably. In contrast, a value type can be a fundamental C++ type, such as int or a class type, such as std::string. If a value consists of other values then is is called a composite value and its type — a composite value type. Otherwise the the value is called simple value and its type — a simple value type. Note that the distinction between simple and composite values is conceptual rather than representational. For example, std::string is a simple value type because conceptually string is a single value even though the representation of the string class may contain several data member each of which would be considered a value. In fact, the same value type can be viewed (and mapped) as both simple and composite by different applications.

Seeing how all these concepts map to the relational model will hopefully make these distinctions more clear. In a relational database an object type is mapped to a table and a value type is mapped to one or more columns. A simple value type is mapped to a single column while a composite value type is mapped to several columns. Conversly, an object is stored as a row in this table and a value is stored as one or more cells in this row. A simple value is stored in a single cell while a composite value occupies several cells.

Going back to the distinction beetween simple and composite values, consider a date type which has three integer data members: year, month, and day. In one application it can be conidered a composite value and each member will get its own column in the relational database. In another application it can considered as a simple value and stored a single column as a number of day from some predefined date.

Until now, we have been using the term persistent class to refer to object classes. We will continue to do so even though a value type can also be a class. The reason for this assimetry is the subordinate nature of value types when it comes to database operations. Remember that values are never stored directly but rather as part of an object that contains them. As a result, when we say that we want to make a C++ class persistent or persist an instance of a class in the database, we invariably refer to an object class rather than a value class.

To make a C++ class a persistent object class we need to declare it as such using the db object pragma:

    #pragma db object
    class person
    {
      ...
    };

The other pargma that we need to use is the db id which designates one of the data members as an object id:

    #pragma db object
    class person
    {

    private:
      #pragma db id
      unsigned long id_;
    };

These two pragmas are the minimum required to declare a persistent class. Other pragmas can be used to fine-tune the persistence-related properties of a class and its members.

You may be wondering whether we aslo have to do declare value types as persistent. We don't need to do anything special for simple value types such as int or std::string since the ODB compiler knows how to map them to the database system types and how to convert between the two. On the other hand, if a simple value is unknown to the ODB compiler then you will need to provide the mapping to the database system type and, possibly, the code to convert between the two. For more information on this see @@ Ref Custom value types/pragma value type. Composite value types are not yet supported by ODB.

Normally, you would use object types to model real-world entities, things that have their own identity. For example, in the previous chapter we created a person class to model a person which is a real-world enitity. Name and age, which we used as data members in our person class are clearly values. It is hard to think of age 31 or name "Joe" as having their own identity.

A good test to determine whether something is an object or a value is to consider if other objects might reference it. A person is clearly an object because it can be refered to by other object's such as a spouce, an employer, or a bank. On the other hand, a person's age or name is not something that other objects would normally refer to.

Also, when an object represents a real entity, it is easy to choose a suitable object identifier. For example, for a person there is an established notion of an identifier (SSN, student id, passport number, etc). Another alternative is to use person't email address as an identifier.

Note, however, that these are only guidelines. There could be goot reasons to make something that would normally be a value an object. Consider, for example, a database that stores a vast number of people. Many of the person objects in this database have the same names and surnames and the overhead of repeating them in every object may negatively affect the performance. In this case we could make first name and last name each an object and only store references to these objects in the person class.

An instance of a persistent class can be in one of two states: transient and persistent. A transient instance only has a representation in the applciation's memory and will ceas to exist when the application terminates unless it is explicitly made persistent. A persistent instance has a representation in both the application's memory and the database. A persistent instance will remain even after the application terminates unless and until it is explicitly deleted from the database. In other words, a transient instance of a persistent class behaves just like an instance of any ordinary C++ class.

3.2 Transactions and Concurrency

A transaction is an atomic, consistent, isolated and durable (ACID) unit of work. All database operations can only be performed within a transaction and each thread of execution in an application can have only one active transaction at a time.

By atomicity we mean that when it comes to making changes to the database state within a transaction, either all the changes succeed or none at all. Consider, for example, a transaction that transfers funds between two objects representing bank accounts. If the debit function on the first object succeeds but the credit function on the second fails, the transaction is rolled back and the database state of the first object remains unchanged.

By consistency we mean that a transaction must take all the objects stored in the database from one consistent state to another. For example, if a bank account object must reference a person object as its owner and we forget to set this reference before making the object persistent, the transaction will be rolled back and the database will remain unchanged.

By isolation we mean that the changes made to the database state during a transaction are only visible inside this transaction until and unless it is commited. Using the above example with bank transfer, the results of the debit operation performed on the first object is not visible to other transactions until the credit operation is successfully completed and the transaction is commited.

By durability we mean that once the transaction is committed, the changes that it made to the database state are permanent and will survive failures such as an application crash. From now the only way to alter this state is to execute and commit another transaction.

Note that all of the above guarantees only apply to the object's state in the database as opposed to the object's state in the application's memory. It is possible to roll a transaction back but still have changes from this transaction in the application's memory. An easy way to avoid this potentiall inconsistency is to instantiate persistent objects withing the transaction's scope. Consider, for example, this two implementations of the same transaction:

void
update_age (database& db, person& p)
{
  transaction t (db.begin_transaction ());

  p.age (p.age () + 1);
  db.store (p);

  t.commit ()
}

In the above implementation, if the store() call fails and the transaction is rolled back, the state of the person object in the database and the state of the same object in the application's memory will differ. Now consider an alternative implementation which only instantiates the person object for the duration of the transaction:

void
update_age (database& db, unsigned long id)
{
  transaction t (db.begin_transaction ());

  auto_ptr<person> p (db.load<person> (id));
  p.age (p.age () + 1);
  db.store (p);

  t.commit ()
}

Of course, it may be not always be possible to write the application in this style. Oftentimes we need to access and modify application's state of persistent objects out of transactions. In this case it may make sense to try to roll back the changes made to the application state if the transaction was rolled back and the database state remains unchanged. One way to do this is to re-load the object's state from the database:

void
update_age (database& db, person& p)
{
  try
  {
    transaction t (db.begin_transaction ());

    p.age (p.age () + 1);
    db.store (p);

    t.commit ()
  }
  catch (...)
  {
    transaction t (db.begin_transaction ());
    db.load (p.id (), p);
    t.commit ();

    throw;
  }
}

A transaction is started by calling the begin_transaction() database function. The returned transaction handle is stored in an instance of the odb::transaction class which has the following interface:

namespace odb
{
  class transaction
  {
  public:
    typedef odb::database database_type;

    void
    commit ();

    void
    rollback ();

    database_type&
    database ();

    static transaction&
    current ();

    static bool
    has_current ();
  };
}

The commit() function commits a transaction and rollback() rolls it back. Unless the transaction has been finalized, (explicitly commited or rolled back), the destructor of the odb::transaction class will automatically roll it back when the transaction instance goes out of scope.

The database() function returns the database this transaction is working on. The current() static function returns the currently active transaction for this thread. If there is no active transaction, this function throws the odb::not_in_transaction exception. You can check whether there is a transaction in effect using the has_current() static function.

If two or more transaction access or modify more than one object and are executed concurrently by different applications or by different threads within the same application, then it is possible that these transactions will try to access objects in an incompatible order and deadlock. The canonical example of a deadlock are two transactions in which the first has modified object1 and is waiting for the second transaction to commit its changes to object2 so that it can update object2. At the same time the second transaction has modified object2 and is waiting for the first transaction to commit its changes to object1 because it also needs to modify object1. As a result none of the two transactions can complete.

The database system detects such situations and automatically aborts the waiting operation in one of the deadlocked transactions. In ODB this translates to the odb::deadlock exception being thrown from one of the database functions. You would normally handle a deadlock by restarting the transaction, for example:

for (;;)
{
  try
  {
    transaction t (db.begin_transaction ());

    ...

    t.commit ()
    break;
  }
  catch (const odb::deadlock&)
  {
    continue;
  }
}